Machine Learning model for the classification of individuals at risk of dementia type Alzheimer from multimodal databases of EEG and clinical information

Abstract

ABSTRACT : Alzheimer's disease (AD) poses a significant challenge in Colombia due to the growing aging population. Detecting early signs of cognitive alterations is crucial, and electroencephalography (EEG) has emerged as a valuable tool for studying AD-related brain activity. However, challenges exist in obtaining comparable and high-quality EEG recordings. Standardized data preprocessing pipelines and harmonization efforts, such as the Brain Imaging Data Structure (BIDS) format, play a vital role in facilitating data integration and sharing. The project focused on organizing multi-site EEG data using the EEG-BIDS framework, promoting localization, accessibility, and interoperability. Open-access databases were utilized to investigate the generalizability of EEG and machine learning (ML) analysis, highlighting the need for data standardization and harmonization. A processing pipeline (Sovaharmony) with normalization and harmonization stages enabled the integration of diverse cohorts (datasets) and optimization of information extraction. Machine learning models were employed for AD risk classification using non-invasive EEG biomarkers. Harmonization of data from multiple cohorts was crucial for increasing sample size, improving statistical power, and identifying consistent features or biomarkers across cohorts. The project aimed to develop a robust and generalizable machine learning model by harmonizing cohorts using a larger and more diverse dataset and thereby improving accuracy. This project made significant contributions to dementia research by developing a comprehensive approach for data acquisition, processing, harmonization, and machine learning-based risk classification using EEG technology. The standardized pipelines, data harmonization, and machine learning techniques were emphasized as critical components in advancing AD research and maximizing the value of EEG data. Further research should focus on replicating the findings on larger cohorts, using techniques like the introduced in the current project, and exploring the application of machine learning models to other non-invasive biomarkers, ultimately validating the accuracy and reliability of AD classification.